Evolving Agents, director plugin
Evolving Agents (director)
The director plugin for Claude Code compiles CLAUDE.md and memory into linked team rules and puts them before the agent's plan.
- 456 ★ on the repository
- Last commit
- Apache-2.0
Extended access
Why this level
- Reads `CLAUDE.md`, `AGENTS.md` and project memory and spends money building rules
- Routes tasks to different models and machines, so a wrong confidentiality label sends data to the wrong place
Install
- Works in
In your terminal, with SkillFoxx CLI
npx skillfoxx add plugins/evolving-agents-directorDetects the agents on your machine, shows the access level and pins the version.
Other ways to install
Access and checks
How we reviewExtended accessThe tool runs code or makes network calls and only reads external data.
- Reads `CLAUDE.md`, `AGENTS.md` and project memory and spends money building rules
- Routes tasks to different models and machines, so a wrong confidentiality label sends data to the wrong place
This is third-party code. Review the repository files before installing.
What it does
The director remembers how a team works: decisions, conventions and hard-won methods become a small memory of linked When / Do / Because rules. Before a plan it finds the applicable rules and puts them in front of the agent, and the plan names the rule each step followed. The `/director:compile` command turns `CLAUDE.md`, `AGENTS.md` and memory into rules, asking first. After a task `/director:reflect` turns your correction into a proposed lesson that is accepted or rejected with Diff, Accept and Reject buttons in a band above the prompt. Tasks marked confidential go only to cleared machines, never to cloud models.
Who it is for. Teams that have worked with Claude Code for a while, have built up `CLAUDE.md` and conventions, and want the agent to follow them when planning instead of guessing.
Good fit when
- The agent again picked a metric definition other than the one the team fixed, and the plan must tie to the accepted decision.
- After correcting a result, a lesson should be drafted and kept only on explicit agreement.
- Work materials are confidential and have to go only to a local model or your own server.
- A project has many `CLAUDE.md`, `AGENTS.md` and notes that need to become a checkable rule set.
Not a fit when
- The project is small and has few rules: building the memory will not pay off.
- The agent is not Claude Code: the plugin runs only there today.
- General reasoning gains from the model are expected: by the authors' own assessment the director adds team knowledge, not general good practice.
Example request
Run `/director:compile` for this project, then plan the schema migration and show which team rule each step uses.Limitations
The plugin runs only inside Claude Code. Building memory with `/director:compile` costs money: the README estimates about US$5 for 60 files. The authors state plainly that with the strongest models it adds no general practice and pays off only on your team's own knowledge. Measurements come from small situation sets, 14 for planning and 8 for learning. The repository also holds the earlier 2025 Evolving Agents Toolkit, and only `packages/director/` is active.
How to disable. Remove the `director` plugin with the Claude Code plugin manager commands. The memory sits on your machine as plain Markdown and is deleted by hand.
FAQ
What does `/director:compile` do?
It turns your `CLAUDE.md`, `AGENTS.md` and memory into linked rules. It asks first, and the README estimates about US$5 for 60 files.
How does learning from corrections work?
After a task `/director:reflect` turns your correction into a proposed lesson. A band above the prompt shows Diff, Accept and Reject, and nothing is saved without your agreement.
Where is the memory stored?
On your machine as plain Markdown. Every accepted lesson is a versioned change you can compare and undo.
What about confidential data?
A task is marked confidential and then goes only to cleared machines: a local model or an office server. The README says code enforces this rule rather than asking the model.
Does it work in Codex or Cursor?
Only in Claude Code today. Running on other agent platforms is named a future plan in the README.
Discussion
No questions or reviews yet.
Related
PluginsA template for Claude Code projects: parallel agent swarms, workflow skills, safety hooks and rules in one install.
Multi-agent orchestration for Claude Code: teams, roles, autopilot and task routing across skills and subagents
A self-improvement plugin for Claude Code, Codex and OpenCode: turns corrections into rules the agent follows in later sessions
Agentic Substrate
Agentic Substrate (claude-user-memory)
A short reversible rule set for an agent: research the unclear, plan proportionally, verify the outcome and keep checked knowledge
Workflows for Claude Code: /do, /omo, /bmad-pilot and more commands, with Codex, Gemini or OpenCode writing the code through codeagent-wrapper
AI Builder Club Skills
AI Builder Club — Skills
A Claude Code plugin marketplace to make a repo agent-ready and run compounding agent loops
README in short
The README explains three mechanics: rule injection before a plan, learning from your corrections and task routing that protects confidential data. Measurements live in a separate `DIRECTOR-EVIDENCE.md`.